The TRAIP handbook

TRAIP is an open catalogue that gathers AI projects built in Türkiye in one place. The team that built a project enters the record themselves; the demo, the technical documentation and the credits are all in their own words.

The name is three pieces: TR (Türkiye), AI, and P (platform) — TRAIP. In full: Türkiye AI Platform.

Records go live after a human moderator approves them. Moderation is not a quality contest: it exists so the catalogue is made of real work, not to rank projects against one another.

Metrics and performance claims are self-reported by the teams. TRAIP does not verify them independently, and they are labelled that way everywhere they appear.

You need an account to add a project. E-mail and a password are enough to register; you can fill in the profile later. Your profile page gathers the projects whose credits name you.

Privacy settings live under Dashboard → Settings. Your e-mail address is hidden by default: addresses published in the open get scraped and come back to you as spam. Visitors can write to you without seeing it.

• Inbox — turning it off removes the contact form from your project pages.
• Profile indexing — turned off, your profile still opens by link but stays out of search results and the sitemap.
• E-mail notifications — for messages, moderation decisions and team invites.

A project is saved as a draft first; nobody else sees it. Edit it as long as you like and send it for review when it is ready. The required fields are the name, the tagline, the overview, the field and the stage.

Name

The project's name. Its URL slug is derived from this and stays fixed once published.

Tagline

What it does, in one line. This is the line shown on cards and in search results; at least 20 characters.

Overview

What the project is, who it is for and what it solves. Markdown is accepted.

How it works

The technical account: model, data, method. Optional, but this is the field that makes the catalogue worth reading.

Limitations

What it cannot do, where it fails, ethical notes. Filling this in does not weaken the record; it makes it trustworthy.

Field

Language, vision, speech, health, agriculture, robotics, education, finance or other. The catalogue's filters run on this.

Stage

Idea, prototype, beta or live. Claiming a later stage than the truth serves nobody.

Links

Demo, the project's own site, source repository (GitHub, GitLab), container image (Docker Hub, GHCR), video, documentation and paper addresses. All optional; each must start with https://.

Logo

The project's own mark. Separate from the cover: a cover is a wide image, a logo is a small square that has to stay legible down to 32 pixels.

Tags

Free-form keywords, for what the field and stage cannot capture: the model used, the dataset, how it is deployed.

Hide from search engines

For pre-launch work, or a project still in academic review. The page stays in the catalogue; it is only left unindexed.

PNG, JPEG and WebP are accepted, up to 4 MB. The type is determined from the file's leading bytes, not its extension — a renamed file does not pass. SVG is refused because it can carry script.

Images live in object storage but are served through the application, so that indexed addresses do not break when the storage provider changes.

You can add anyone to the credits: someone with an account, an advisor without one, an outside contributor. For each you write a name and a role — "Model", "Data and labelling", "Advisor".

When you add someone who has an account, an invite goes out and they do not appear in the credits until they accept. The reason is simple: nobody wants their name attached to a project without their knowledge.

• Documents — links to a README, an API reference, a PDF, a dataset or a notebook. Each gets a title and, if you like, a short note such as "8 min" or "v0.4".
• Releases — a version number, a date and what changed in it. This is where a project shows movement rather than stillness.
• Metrics — a label, a value and optionally a comparison context ("human: ~14 min"). Everywhere they appear, they are marked as self-reported.

Draft

Only you see it. There is no time limit.

In review

Waiting in the moderation queue. You can keep editing while it waits.

Changes requested

The moderator has left a note. Fix it and send it again; you do not start over.

Published

Visible in the catalogue and open to search engines and agents. You can withdraw it at any time.

Only a human moderator decides to publish. Neither the API nor MCP has a publishing endpoint — even an agent acting for you can only send a project for review, never put it live.

Anyone with an account may rate a published project once, and add a comment if they want to. The average appears on the project page and on catalogue cards.

• An account is required and one person counts once. An average nobody can push with repeated votes measures what it claims to.
• An owner cannot rate their own project.
• You can change your rating later; the average corrects for it.
• A moderator can hide a review. The row stays, but its rating leaves the average — a review nobody can read should not keep steering the score.

Every published project page carries a contact form. A visitor writes, you read it in your dashboard, and you get an e-mail notification if you want one. Your address never appears on the page at any point.

You create a token under Dashboard → API tokens. It is shown only at the moment of creation; only its hash is kept in the database. If you lose it, it cannot be recovered — you issue a new one.

• A token acts with its owner's permissions and reaches only their projects.
• Anything belonging to someone else answers 404, not 403 — not even the existence of another account's work leaks.
• Validation matches the web form exactly; a record sent through the API cannot be in a state the form would reject.
• You can revoke a token from the settings page; it stops working immediately.

If your assistant speaks the Model Context Protocol, the same API is available as tools. The full endpoint list, examples and the MCP setup are in the developer documentation.

The interface is offered in Turkish, English, French, German, Spanish, Chinese and Arabic. Turkish sits at the root address; the others take a prefix. Arabic lays out right to left.

Only the Turkish and English versions are offered to search engines. The other five translate the interface but not the projects; indexing them all would publish five near-duplicates of every page. Those pages remain reachable by link.

Published records are open and also served in machine-readable form. The whole catalogue can be fetched in one request; each project's record can be read from its own address.

/catalog.json
/llms.txt
/proje/<kisa-ad>/record.json
/documentation.ai.md

Drafts and records in review appear on no public endpoint. E-mail addresses are published nowhere. Profiles you have closed stay out of the sitemap.

What the catalogue attests is that a team recorded their work with these descriptions. It does not attest that a tool works, that its metrics are accurate, or that any listing covers all the work done in Türkiye.